Fig. 1: Architecture of OPUS-DSD.
From: OPUS-DSD: deep structural disentanglement for cryo-EM single-particle analysis
a, Schematic diagram of the architecture. Pose refers to the projection direction of input with respect to the consensus model. For simplicity, the standard Gaussian prior for latent distribution as well as the smoothness and sparseness priors for the 3D volume are omitted in this chart. b, Architecture of the encoder of OPUS-DSD. This diagram shows the encoder that translates a 2D cryo-EM image into the latent encoding. The top row denotes the dimensions of the intermediate tensors. The arrow links the input and output of the operation above. FC, fully connected layer; Conv3D, 3D convolution; ST, spatial transformer (which back-projects the 2D image to a 3D volume). The number of channels of the convolution kernel can be derived from the dimensions of its input and output. The ellipsis represents the repeating of the preceding operation until the tensor reaches the output dimension. All convolutions and fully connected layers except the last one had LeakyReLU33 (leaky rectified linear unit) non-linearity with a negative slope of 0.2. c, Architecture of the decoder of OPUS-DSD. This diagram shows the decoder that translates the latent encoding z into a reconstructed 2D projection. Conv3DT, 3D transposed convolution; ST, spatial transformer, which here renders the 3D volume into the 2D image of desired resolution. All transposed convolutions except the last one and fully connected layers have LeakyReLU non-linearity with a negative slope of 0.2.
